Paris School of International Affairs - Sciences Po

Success would require the Dialogue to demonstrate, from its first session, that inclusive global deliberation on AI can produce something more than agreed language, and that the UN system is capable of hosting that conversation without reducing it to the lowest common denominator of diplomatic consensus. A first marker of success would be substantive engagement with the asymmetry at the heart of the AI governance challenge: the countries most exposed to AI's risks and least able to capture its benefits are also those with the least capacity to shape the governance frameworks that will affect them. A Dialogue that produces a co-chair summary treating all Member States as equally situated would have missed its own point. Success means naming that gap directly and identifying concrete referrals — to UNESCO, ITU, UNDP, or other operational bodies — capable of acting on it. A second marker would be the quality of the connection between the Scientific Panel's evidence and the deliberations of Member States. Success means that connection is visible and traceable in the outcome. Not that governments agreed on everything, but that the deliberations were genuinely informed by evidence rather than by competing political positions alone. Third, and looking forward: the second Dialogue is already scheduled for 2027. Success in July 2026 would mean arriving in New York with specific, unresolved questions to continue, a live agenda. A first session that opens productive disagreement will already mean some form of success.

• AI capacity-building
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Transparency, accountability, and human oversight

These three areas are selected not as independent priorities but as components of a single, connected challenge. Meaningful transparency and human oversight of AI systems - however well designed in law or institutional architecture - depend on the existence of humans with sufficient understanding to exercise them. A regulator who cannot interrogate a model's outputs, an educator who cannot assess the pedagogical implications of an AI tool, a civil servant who cannot distinguish a system's stated confidence from its actual reliability: none of these actors can provide the oversight that governance frameworks mandate, regardless of the formal authority they hold. Transparency mechanisms without literacy to interpret them are procedural fictions. This is why I do not see capacity-building as a secondary or developmental concern, but rather as the foundational one. And capacity-building in AI cannot be reduced to infrastructure and connectivity, important as those are. It must include the human dimension: training educators, equipping policymakers, reforming the curricula of the institutions that prepare the next generation of public officials and professionals. The social, economic, ethical, cultural and linguistic implications of AI provide the substantive terrain on which both of the above operate. AI systems are not culturally neutral. They reflect the priorities, languages, and values embedded in their training data and design choices. Governance that does not grapple seriously with those asymmetries will produce frameworks that entrench existing inequalities rather than correct them. These three areas, taken together, describe the minimum conditions for AI governance that is genuinely functional, and not merely formal.

Two issues deserve explicit attention that the current thematic framework does not adequately capture. The first is AI literacy as a structural precondition for governance. The seven themes address capacity-building, oversight, transparency, and human rights. None of them names the foundational obstacle that runs across all of them: the acute deficit of AI understanding among the people responsible for governing, regulating, teaching, and overseeing AI systems. A transparency mechanism interpreted by officials who cannot read its outputs is a procedural fiction. A human oversight requirement exercised by regulators who cannot interrogate a model's behaviour is a formal mandate without substance. AI literacy is not one competency among many - it is the condition without which the other six themes cannot be implemented meaningfully. It cuts across every cluster of this Dialogue and should be named explicitly as a cross-cutting priority, with dedicated attention to the institutions responsible for building it: schools, universities, teacher training programmes, and the professional formation systems that prepare public officials and international civil servants. The second is the specific situation of children and adolescents. AI and human rights is a broad and necessary category; the situation of minors within it is sufficiently distinct to warrant separate treatment. Children are encountering AI systems during the years in which their cognitive habits, critical faculties, and relationship to knowledge are being formed. The developmental implications of that exposure are not the same as those for adults, and they are not captured by general human rights or transparency frameworks. International governance has a long tradition of treating the rights and protection of minors as a distinct domain. AI governance should do the same, and this Dialogue is the right place to establish that precedent.

The education sector illustrates the governance gap with particular clarity, and at a scale that makes it one of the most urgent sites of intervention. AI tools are already present in classrooms, in homework workflows, in admissions processes, in learning management systems, and in the assessment practices of institutions at every level. Their integration is largely market-driven and rarely pedagogically deliberate. Decisions about the role of AI in students' intellectual development are being made by default — by platform providers and procurement processes — rather than by educators equipped to make those decisions intentionally. This should not be seen as a future risk, but rather as the current situation in most education systems, in high-income and low-income countries alike, with variations in the tools involved but not in the structural dynamic. The challenge is particularly acute in higher education, and most consequentially in the institutions that train future policymakers, lawyers, economists, diplomats, and public administrators. These are the professionals who will be expected to regulate, oversee, and govern AI systems. If their formation does not equip them to engage critically with those systems, the governance frameworks they will be asked to implement remain structurally hollow — well-designed on paper, unenforceable in practice. The opportunity is equally clear. Education is the sector where investment in AI literacy produces the longest-lasting and most widely distributed returns. The pedagogical challenge and the governance challenge are the same challenge viewed from different ends of the same pipeline. Addressing one without the other produces neither. I believe that there may be a high degree of awareness of the problem. But that is not enough. What is lacking is the institutional commitment to treat it as the structural priority it is.

The Dialogue's most realistic and valuable contribution to international cooperation is the construction of a shared vocabulary and a common evidence base from which agreement becomes eventually possible over time. AI governance is currently fragmented across regional frameworks, national legislation, and sectoral initiatives that reflect different values, different risk assessments, and different institutional capacities. That fragmentation is of course not a problem of bad faith. It is a problem of insufficient shared understanding of what AI systems actually do, of what governance mechanisms have worked and which have not, and of what the distributional consequences of different regulatory choices look like across different national contexts. The Dialogue, supported by the Independent Scientific Panel, is structurally positioned to address precisely that deficit. This means the Dialogue's value is cumulative and iterative rather than immediate. We all know well that a single session will not produce convergence on any contested questions. But a recurring, inclusive forum, bringing a large number of States into the same conversation, well grounded in independent evidence rather than industry framing, can progressively narrow the space of legitimate disagreement and build the mutual understanding that formal cooperation requires. The condition for this to work is that the Dialogue maintains genuine inclusivity in substance, and not only in participation. Countries with limited AI capacity must be able to shape the agenda. The scientific evidence presented must be accessible across languages and levels of technical familiarity. And the outcomes must feed visibly into operational mechanisms (in UNESCO, ITU, and elsewhere) that can translate deliberation into action. This approach may well help the Dialogue become somethimng else than a well-attended conversation that changes nothing.

The landscape of existing AI governance initiatives is active but fragmented. The Dialogue's added value lies precisely in its capacity to connect efforts that currently operate in parallel without sufficient coordination. Several existing mechanisms deserve explicit recognition. UNESCO's 2021 Recommendation on the Ethics of AI remains the only globally adopted ethical framework for AI, and its implementation observatory provides a practical monitoring infrastructure that the Dialogue should build upon rather than duplicate. The ITU's AI for Good platform and its international standards work anchor the technical dimension. The OECD's AI Policy Observatory offers a comparative policy database that is genuinely useful for evidence-based deliberation, though its membership does not reflect the full diversity of UN Member States. The Council of Europe's AI Convention represents the most advanced binding instrument to date, with relevance beyond its immediate signatories. At the research and civil society level, institutions such as the Ada Lovelace Institute, AlgorithmWatch, and the Future of Life Institute contribute independent analysis that is not captured by intergovernmental processes. What most of these initiatives share is a limited reach: they reflect the priorities and capacities of the countries and institutions that created them. The Dialogue's added value is not to produce a better version of any one of them. It is to create the conditions under which their outputs can be stress-tested against a genuinely global range of perspectives — including those of countries that had no role in shaping the frameworks now being presented to them as international standards. That is a modest but real contribution. It will not resolve the fragmentation overnight. But it can begin to build the connective tissue that coherent international cooperation on AI governance will eventually require.

Multi-stakeholder participation in global AI governance forums has a structural weakness that format design rarely addresses directly: the actors with the most resources and the most at stake commercially are systematically better positioned to participate than those with the most to lose socially. Industry representatives arrive with prepared positions, dedicated staff, and established relationships with delegations. Civil society organisations, academic researchers, and representatives from lower-income countries arrive, when they arrive at all, with none of those advantages. Formal equality of access does not produce substantive equality of influence. The Dialogue's format should be designed with that asymmetry explicitly in mind. This means, at minimum: advance publication of substantive background documents in all UN languages, with sufficient lead time for preparation; structured slots for civil society and academic contributions that are not residual. Not the last fifteen minutes of a session, but integrated into the deliberative architecture; and a clear mechanism by which inputs from the written consultation process, including submissions from individuals and smaller organisations, are visibly reflected in the framing of discussions rather than archived unread. On the question of academic and scientific contribution specifically: the relationship between the Independent Scientific Panel and the deliberative sessions should be interactive, not presentational. Governments and stakeholders should not simply receive the evidence, but also be able to engage with it, contest it, and request clarification. Anything that moves forward from the classical rigidity of a panel that presents and a room that listens would be progress. This forum should pay tribute to its name: Dialogue.

The most predictable answer to this question names the Global South, indigenous communities, and civil society from lower-income countries. That answer is correct, and the structural recommendations made elsewhere in this submission — genuine linguistic accessibility, substantive rather than formal inclusion — apply directly to those communities. They do not need to be repeated here. Still, two less commonly named gaps deserve attention. The first is the absence of educators and education professionals as a distinct voice in AI governance deliberations. These are the people closest to the effects of AI on human development, learning, and the formation of critical judgment. They are not organised as a governance constituency, they hold no dedicated seats in multi-stakeholder forums, and their practical, daily knowledge of how AI is reshaping the conditions of learning is not systematically collected or fed into policy processes. A governance conversation about AI's social and cultural implications that does not include the people responsible for transmitting culture and developing social capacity is missing one of its most important interlocutors. The second is the quality of youth participation. Young people appear in global AI governance forums, but typically in a ceremonial capacity: a designated speaker, a symbolic presence, an affirmation that the future generation has been acknowledged. What is structurally absent is the substantive engagement of the generation that is already navigating AI in ways most senior policymakers are not, and that will live longest with the consequences of decisions made today. Youth participation should be designed to introduce analytical perspectives, not to add colour to an agenda shaped entirely by others. Both gaps point to the same underlying risk of failure: governance processes that mistake the presence of a category for the inclusion of a voice.

Format choices in global governance forums tend to reflect the preferences and habits of the institutions convening them rather than the conditions most likely to produce genuine exchange. The result is a familiar architecture: plenary statements, moderated panels, side events — formats optimised for the transmission of prepared positions rather than for the kind of mutual adjustment that dialogue actually requires. The most useful innovation would not be a new format but a different principle of design: sessions should be structured around unresolved questions rather than around topics. The difference matters. A session on "AI and human rights" invites statements. A session organised around a specific, genuinely contested question — one to which governments and stakeholders hold meaningfully different answers — creates the conditions for substantive exchange. The co-chairs and secretariat are best placed to identify which questions those are, based on the consultation inputs received. I would simply urge that the effort be made. One concrete suggestion: structured small-group exchanges between government representatives and non-state actors — not side events, but integrated into the main programme — where the explicit purpose is for each to engage with the other's constraints and priorities rather than to present their own. The format already exists in various negotiating contexts. Its application here would be a modest but real departure from the standard conference model. Beyond that, I would not dare to instruct experienced multilateral practitioners on meeting design. The principle is what matters: format should serve dialogue, not substitute for it.

The most instructive lesson from existing AI governance efforts is not contained in any single framework or platform. It emerges from the pattern of what has worked and what has not. The governance approaches that have produced measurable change share a common characteristic: they embed competence-building into the regulatory design itself, rather than treating regulation and capacity as parallel but separate tracks. Effective oversight requirements are accompanied by training for the people expected to exercise them. Implementation mechanisms are designed around the actual capabilities of the institutions responsible for applying them. The instrument and the literacy required to operate it are conceived together, not sequentially. The approaches that have produced the least change share a different characteristic: they assumed the competence that was needed rather than building it. The proliferation of AI ethics principles and voluntary frameworks over the past decade is the clearest example. It generated extensive documentation, broad nominal endorsement, and negligible behavioural change. The reasons are consistent across cases: principles without enforcement, accountability without the capacity to verify compliance, and governance actors who lacked the analytical grounding to translate commitments into decisions. This pattern has a direct implication for the Dialogue. When evaluating existing initiatives to build upon, the relevant question should be which ones were designed with an honest assessment of the capacities available to implement them, and which ones were designed for a world of competent, well-resourced governance actors that does not yet exist in most of the countries that signed them. Closing that gap between the governance frameworks the world has written and the institutional capacity needed to make them real is, in this submission's view, the central task of international AI governance in the years ahead.